Medical Image Training Data Generation via Motion Weighted Blur Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for medical image analysis fail to accurately simulate non-uniform motion velocities and complex motion patterns, such as acceleration and deceleration, which are prevalent in medical imaging.
Innovation Solution
A system and method for generating training data that involves modifying a blur convolution kernel with pixels oriented in the direction of movement, using motion weighting factors to simulate realistic motion blur effects, including acceleration and deceleration, in medical images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If motion blur is simulated using traditional methods with uniform velocity, then the training data generation is simple, but the realism of motion patterns is insufficient
Solution Approach 1:
The patent applies dynamics by transitioning from static uniform velocity to dynamic non-uniform velocity patterns. The system introduces temporal variations in motion velocity, acceleration, and deceleration phases to simulate realistic physiological motion. This is achieved by modifying the blur convolution kernel to incorporate time-varying velocity profiles that match cardiac and respiratory motion patterns, thereby improving motion realism without excessive complexity.
Solution Approach 2:
The patent changes key motion parameters including velocity, acceleration, and temporal duration to create realistic motion blur. By varying these parameters across different phases of the cardiac and respiratory cycles, the system generates more authentic training data. The blur kernel is constructed with parameterized velocity profiles that can be adjusted to match different physiological conditions and imaging scenarios.
2Measurement precision
If complex motion patterns with acceleration and deceleration are simulated, then the accuracy of motion recognition improves, but the data generation process becomes more complex
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing blur convolution kernels for various motion patterns before actual training data generation. The system creates a library of kernels representing different velocity, acceleration, and deceleration profiles that can be directly applied to sharp images. This pre-computation approach simplifies the data generation process while maintaining high accuracy in motion pattern simulation.
Solution Approach 2:
The patent uses copying by generating synthetic motion blur through convolution of sharp images with pre-computed kernels. Instead of capturing real motion-blurred images, the system creates accurate copies by applying mathematical convolution operations. This approach preserves the exact motion characteristics encoded in the kernels while avoiding the complexity of real-world data collection and annotation.
3Reliability
If motion weighting factors are used to simulate non-uniform motion, then the physiological accuracy of training data improves, but the computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the motion simulation into distinct temporal phases: acceleration phase, constant velocity phase, and deceleration phase. Each phase is represented by separate weighting factors applied to the blur kernel. This segmented approach accurately captures physiological motion patterns while keeping computational complexity manageable through modular processing of each motion phase.
Solution Approach 2:
The patent uses partial action by applying motion weighting factors selectively to different regions and time periods of the imaging process. Rather than uniformly applying complex motion models throughout, the system applies weighting factors only where and when they are most needed to capture physiological motion, reducing unnecessary computational overhead while maintaining accuracy in critical regions.
Data Source
AI summary
The present invention relates to training data sets and a system and method for generating training images especially those which are medical images. Especially disclosed is a method of training a machine learning model to recognize movement of a body part in an acquired medical image The machine learning model is trained by varying/modifying a blur convolution kernel constructed with pixels oriented in a direction of the movement; the method including determining at least one motion weighting factor corresponding to a motion time period when the body part is moving during acquisition of the medical image, and using the motion weighting factor to vary/modify the blur convolution kernel.